When we are faced with challenging image classification tasks, we often explain our reasoning by dissecting the image, and pointing out prototypical aspects of one class or another. The mounting evidence for each of the classes helps us make our final decision. In this work, we introduce a deep network architecture -- prototypical part network (ProtoPNet), that reasons in a similar way: the network dissects the image by finding prototypical parts, and combines evidence from the prototypes to make a final classification. The model thus reasons in a way that is qualitatively similar to the way ornithologists, physicians, and others would explain to people on how to solve challenging image classification tasks. The network uses only image-level labels for training without any annotations for parts of images. We demonstrate our method on the CUB-200-2011 dataset and the Stanford Cars dataset. Our experiments show that ProtoPNet can achieve comparable accuracy with its analogous non-interpretable counterpart, and when several ProtoPNets are combined into a larger network, it can achieve an accuracy that is on par with some of the best-performing deep models. Moreover, ProtoPNet provides a level of interpretability that is absent in other interpretable deep models.

We propose a novel technique that allows one to conserve energy using the time integration scheme of one's choice. Traditionally, the time integration methods that deal with energy conservation, such as symplectic, geometric, and variational integrators, have aimed to include damping in a manner independent of the size of the time step, stating that this gives more control over the look and feel of the simulation. Generally speaking, damping adds to the overall aesthetics and appeal of a numerical simulation, especially since it damps out the high frequency oscillations that occur on the level of the discretization mesh. We propose an alternative technique that allows one to use damping as a material parameter to obtain the desired look and feel of a numerical simulation, while still exactly conserving the total energy-in stark contrast to previous methods in which adding damping effects necessarily removes energy from the mesh. This allows, for example, a deformable bouncing ball with aesthetically pleasing damping (and even undergoing collision) to collide with the ground and return to its original height exactly conserving energy, as shown in Fig. 2. Furthermore, since our method works with any time integration scheme, the user can choose their favorite time integration method with regards to aesthetics and simply apply our method as a postprocess to conserve all or as much of the energy as desired.

JFLAP is software for experimenting with formal languages and automata theory. In this Tips and Techniques session we describe the recent changes to JFLAP to make it more usable in an automata theory course.

We describe the results from a two-year study with fourteen universities on presenting formal languages in a more visual, interactive and applied manner using JFLAP. In our results the majority of students felt that having access to JFLAP made learning course concepts easier, made them feel more engaged in the course and made the course more enjoyable. We also describe changes and additions to JFLAP we have made based on feedback from users. These changes include new algorithms such as a CYK parser and a user-controlled parser, and new resources that include a JFLAP online tutorial, a wiki and a listserv.

In this paper we simulate high resolution cloth consisting of up to 2 million triangles which allows us to achieve highly detailed folds and wrinkles. Since the level of detail is also influenced by object collision and self collision, we propose a more accurate model for cloth-object friction. We also propose a robust history-based repulsion/collision framework where repulsions are treated accurately and efficiently on a per time step basis. Distributed memory parallelism is used for both time evolution and collisions and we specifically address Gauss-Seidel ordering of repulsion/collision response. This algorithm is demonstrated by several high resolution and high-fidelity simulations.

We propose a novel solid/fluid coupling method that treats the coupled system in a fully implicit manner making it stable for arbitrary time steps, large density ratios, etc. In contrast to previous work in computer graphics, we derive our method using a simple back-of-the-envelope approach which lumps the solid and fluid momenta together, and which we show exactly conserves the momentum of the coupled system. Notably, our method uses the standard Cartesian fluid discretization and does not require (moving) conforming tetrahedral meshes or ALE frameworks. Furthermore, we use a standard Lagrangian framework for the solid, thus supporting arbitrary solid constitutive models, both implicit and explicit time integration, etc. The method is quite general, working for smoke, water, and multiphase fluids as well as both rigid and deformable solids, and both volumes and thin shells. Rigid shells and cloth are handled automatically without special treatment, and we support fully one-sided discretizations without leaking. Our equations are fully symmetric, allowing for the use of fast solvers, which is a natural result of properly conserving momentum. Finally, for simple explicit time integration of rigid bodies, we show that our equations reduce to form similar to previous work via a single block Gaussian elimination operation, but that this approach scales poorly, i.e. as though four spatial dimensions rather than three.

There has been much research on the benefits of active and collaborative learning and on its use in computer science courses. As classroom technology becomes more prevalent it is natural to develop systems that support the use of these techniques. We have developed such a system as an extension to Classroom Presenter [2], a Tablet PC-based presentation system. In our system students (or groups of students) are equipped with tablet computers and at various points during the lecture, are asked to solve a problem or respond to a question. Students respond by writing their solution on the tablet and submitting it wirelessly to the instructor. The instructor can view all student responses, select one or more to display to the class, and annotate responses with ink as they are being displayed. Student responses can be saved for review after lecture by the instructor or made available electronically to the students. In this paper we describe the system and our initial experiences using the system in two classes (CS1 and Computer Ethics) at the University of San Diego in Fall 2003.

In this paper, we report on an empirical exploration of digital ink and speech usage in lecture presentation. We studied the video archives of five Master's level Computer Science courses to understand how instructors use ink and speech together while lecturing, and to evaluate techniques for analyzing digital ink. Our interest in understanding how ink and speech are used together is to inform the development of future tools for supporting classroom presentation, distance education, and viewing of archived lectures. We want to make it easier to interact with electronic materials and to extract information from them. We want to provide an empirical basis for addressing challenging problems such as automatically generating full text transcripts of lectures, matching speaker audio with slide content, and recognizing the meaning of the instructor's ink. Our results include an evaluation of handwritten word recognition in the lecture domain, an approach for associating attentional marks with content, an analysis of linkage between speech and ink, and an application of recognition techniques to infer speaker actions.